Interpretable machine learning-assisted compositional screening and experimental validation of a silicide-reinforced titanium matrix composite

Developing titanium matrix composites (TiMCs) with balanced strength and ductility remains challenging because compositional optimization involves complex interactions among matrix alloying elements and in-situ reinforcement precursors. In this study, an interpretable machine learning-assisted workflow was developed for compositional screening and experimental validation of Ti alloys and TiMCs. A curated database containing approximately 600 records was established, and four regression models were evaluated for predicting ultimate tensile strength and elongation. Among them, ensemble models, especially Random Forest and XGBoost, showed better predictive capability than linear regression for both properties. SHAP analysis was further employed to examine the relative effects of key elements and to identify data-supported compositional regions associated with favorable strength-ductility combinations within the present dataset. Based on the combined screening results, Ti-6.5Al-4Sn-3Zr-1Mo-0.7Nb-1W-0.3Si-0.06 C was selected as a candidate composition for experimental validation and fabricated by powder metallurgy. The resulting in-situ silicide-reinforced TiMC exhibited a tensile strength of 1201 ± 3.5 MPa with 12.5 ± 0.5% elongation at room temperature, and 464 ± 4.0 MPa with 12.0 ± 0.7% elongation at 750 °C. Microstructural observations revealed refined lamellar features and dispersed Si-rich particles in the composite. The present results demonstrate the feasibility of combining interpretable machine learning with experimental verification for compositional screening and development of TiMCs for elevated-temperature applications.